-
Dictator no longer contacts Hugging Face when you use a model you have already downloaded. Loading a local model was re-checking its files against huggingface.co every time, which meant a request went out on each Assistant Mode activation and whenever a model was reloaded for dictation or meeting notes. No transcript, audio, screen content or prompt was ever sent — but the request itself revealed your IP address, which model you use, and when you used it, and that should not have been leaving your Mac. Both the language-model and Whisper paths now load straight from disk, and only reach the network when a model genuinely needs downloading.
-
Speech recognition and the assistant now work normally on a captive or unreliable network. Loading an already-downloaded model previously had to reach Hugging Face first, so a bad connection could stall or fail it even though every file was already on your Mac.
-
Fixed the Meetings window crashing the app on open after a long meeting — a meeting whose notes never got their final polish pass, shown in a window narrow enough that the sidebar, transcript and Details panes could not all fit, put the window into a layout loop that took Dictator down
-
The Meetings sidebar can no longer be dragged wide enough to squeeze the transcript, and the window keeps room for all three panes
-
The Notes / Transcript / Pad / Coach tabs now scroll sideways when the window is narrow instead of forcing the window wider
-
Dictating into browser comment and chat boxes no longer inserts a stray space at the start of the text — the smart join now verifies the caret snapshot actually describes the focused field, and stands down when the browser reports page text instead
-
The Settings window is rebuilt on native Mac chrome, like System Settings: page titles and controls live in a real toolbar (the Models tabs, the Modes add/back buttons with rename-in-place for the mode name, Dictionary search/sort/add, and the History count and Clear button), the close buttons sit centred in the full-height header, and the sidebar uses the standard material without rendering artifacts
-
Opening a mode in Settings → Modes now slides in like a standard navigation page, with a back button in the toolbar
-
The Settings sidebar now uses the standard translucent macOS sidebar material and extends behind the titlebar, so the window reads like System Settings instead of a flat gray panel (also fixes the sidebar occasionally rendering with the wrong background)
-
Modes are now built from steps. Instead of three fixed passes (Format, Grammar, Structure), each mode is an ordered list of AI steps you can add, remove, reorder, and edit — most modes need just one. New installs get a curated set: Standard (one fast pass), Polished (adds a grammar tidy), Formal (adds tightening), plus the no-AI Quick. Add more anytime from the + gallery in Settings → Modes. Your existing modes carry over unchanged.
-
Dictating a short phrase into an empty field (a fresh chat box, a new note) now capitalises the first letter, since an empty field is clearly the start of a sentence. Short dictations inserted mid-sentence still stay lowercase as before.
-
The assistant result window now shows, per turn, what context was pulled in — whether it read the text around your cursor — with a "Show context" toggle to see the actual captured text. Makes it clear what fed each answer.